AI leaves the lab. We follow what happens next.

Archive September 30, 2026 20:11 UTC
Listen to today's briefing 9:07
Signal: Enforcement, espionage and a frontier model released only to defenders - while the labs' midterm money was already spent in the primaries.
FTC Comes for the Labs a Day After They Self-Policed
Chairman Ferguson opened the investigation weeks ago and the agency is now drafting civil investigative demands to compel testimony from Anthropic, OpenAI and other frontier labs - and from METR, the watchdog that tests them.
CBS News: the FTC confirmed the investigation on Wednesday SecurityWeek: Anthropic flags agent liability as OpenAI is sued over the hack

MI5: 100 UK Academics Fed China's Spies AI Research
MI5 named the China General Technology Research Institute a front for Beijing's ministry of state security, said its purpose is funding research that improves MSS espionage capability, and warned academics they can now be prosecuted under the National Security Act.
BBC: the security minister has written to every vice-chancellor to cut the tie
NEWGoogle's Gemini 4 Ships Ungated - For Defenders Only
Argon tops DeepSWE at 77.9% and lifts the output ceiling from 64K to a million tokens, but it goes first to cyber defenders in Google's Fairwind programme - with the cyber guardrails deliberately switched off.
Ars Technica: Google announces Gemini 4, but you can't use it yet 9to5Google: Argon's 77.9% on DeepSWE against Opus 5.5's 74.2%
NEWOpenAI Names Moonshot in Reasoning-Theft Campaign
OpenAI says people tied to Kimi's maker ran a scaled campaign from the first week of July to reproduce its protected reasoning - no breach, no stolen conversations, just manipulated prompts - and the attribution is OpenAI's own.
Wccftech: the campaign started the week Kimi K3 debuted Unite.AI: OpenAI passed the technique to rivals via the Frontier Model Forum
Ghost Note They Exhumed Garbo to Sell Ball Bearings